Jing Ren, Zhi-Lin Yuan, Zhe Wu, Shi-Ping Yang, Liang-Liang Chen, Jia Zhao, Yu-Ning Cheng, Chen Wang, Xin Gao, Zheng-Yu Jin, Yuan Li, Fu-Ze Cong, Hua-Dan Xue, Yong-Lan He
In this multicenter head-to-head comparison, radiomics did not significantly improve AUC over traditional CT features for OCCC diagnosis. Although sensitivity differences between models were not statistically significant, the integrated model achieved numerically higher and more stable sensitivity across both cohorts, suggesting potential clinical value in reducing missed diagnoses and avoiding ineffective neoadjuvant chemotherapy. These findings warrant further validation in larger prospective cohorts.
BACKGROUND: Platinum-resistant OCCC misdiagnosis risks ineffective chemotherapy. CT morphology is widely used for diagnosis, and our prior single-center study showed radiomics is feasible. Whether radiomics adds value beyond morphology remains unclear. This multicenter study compares CT, radiomics, and integrated models for OCCC diagnosis.
METHODS: 457 patients with epithelial ovarian cancer (training = 280, internal testing = 69, external testing = 108). Two radiologists assessed 10 CT morphological features. From CT, 1,218 radiomic features were ICC-filtered (≥ 0.8) + JMIM selection. Three logistic regression models were built: traditional (clinical + CT morphology), radiomics (selected features, output as rad-score), and integrated (traditional + rad-score). Performance was evaluated using ROC analysis, and rad-score correlation with morphological features was examined.
RESULTS: Of 457 patients, 96 (21%) had OCCC. In internal testing set, the integrated model achieved the highest AUC (0.890) but was not significantly superior to the traditional model (0.840, p = 0.210) or the radiomics model (0.811, p = 0.051); both integrated and traditional models had 100% sensitivity versus 75.0% for radiomics. In external testing, the integrated model maintained an AUC of 0.850, compared with 0.837 for the traditional model (p = 0.645) and 0.809 for the radiomics model (p = 0.102); sensitivity was 81.8%, 77.3%, and 81.8%, respectively. The rad-score significantly correlated with multiple CT morphological features.
CONCLUSIONS: In this multicenter head-to-head comparison, radiomics did not significantly improve AUC over traditional CT features for OCCC diagnosis. Although sensitivity differences between models were not statistically significant, the integrated model achieved numerically higher and more stable sensitivity across both cohorts, suggesting potential clinical value in reducing missed diagnoses and avoiding ineffective neoadjuvant chemotherapy. These findings warrant further validation in larger prospective cohorts.